### Summary
This PR fixes a bug that @as-cii and @osiewicz saw when the on-disk
contents of files changed due to running `git checkout` at the command
line. It caused a buffer's contents to diverge from the file's on disk
contents, but the buffer to show an *unmodified* status.
I've also introduced new APIs on gpui's deterministic executor, which
make it possible to write a test that reliably triggered the bug.
### Details
The bug is triggered by the following sequence of events:
1. A buffer's file changes on disk while the buffer is *unmodified*
2. Zed reloads the new content of the file
3. Before updating the buffer itself, Zed computes a *diff* between the
buffer's current contents, and the newly-loaded contents
4. While this diff is being computed, one of two things happens:
1. the buffer changes on-disk *again*.
2. the user edits the buffer, but undoes the edit, so that the buffer
returns to an unmodified state
The bug itself was caused by a few things:
* The buffer diffing algorithm is pretty slow, because we perform a
character-wise diff
* We previously allowed multiple reload tasks to run concurrently
* When discarding an out-of-date diff, we failed to update parts of the
buffer's state (`saved_fingerprint`) which allow us to recognize that
the buffer's content differs from the file.
It was also difficult to reproduce the problem in tests, because under
deterministic execution, because it was extremely unlikely for other
tasks to make progress *after* a file had been reloaded, but *before*
the disk task has resolved. To help with testing, I introduced a pair of
executor APIs:
`spawn_labeled`, - for spawning a background task with a given *label*
`deprioritize_task` - for forcing tasks with a given label to run
*after* all other concurrent tasks.
I also made the `Model::next_event` test helper method more useful, in
that it no longer runs *until* parked in order to wait for the next
event to occur. It just steps the executor one poll at a time until the
model emits an event.
Release Notes:
- Fixed a bug that caused buffers to report incorrect modified/conflict
status when their buffers changed on disk multiple times in rapid
succession.
This required using mpsc channels to invoke frame callbacks on the
main thread and send the receiver to the platform display link.
Co-Authored-By: Julia Risley <julia@zed.dev>
The entity map needs to be able to distinguish between the case when
the entity_id is waiting to be dropped, and when it is completely gone.
Before 8bc207141, it assumed that entity_ids in dropped_entity_ids could
be re-used. This caused `take_dropped` to error because the slot had
been overwritten. The fix there caused weak handles to allow upgrading
a reference count from 0, which could resurrect items in
`dropped_entity_ids` which caused them to be dropped twice.
We could allow weak items to upgrade from 0, and delete from
dropped_entity_ids, but that seemed more complicated than necessary.
Piotr Osiewicz
Mikayla
Max Brunsfeld
Conrad Irwin
Marshall Bowers
Antonio Scandurra
